#ffmpeg #tensorflow #tesseract #gpu #video

hypetrigger

Perform efficient per-frame operations on streaming video

2 unstable releases

0.2.0 Jan 19, 2023
0.1.0 Jan 12, 2023

#42 in Multimedia

Download history 19/week @ 2023-01-07 34/week @ 2023-01-14 14/week @ 2023-01-21

67 downloads per month

MIT license

63KB
1K SLoC

Hypetrigger ⚡

Perform efficient per-frame operations on streaming video.

Links | hypetrigger.io | Github | crates.io | docs.rs | npm

Architecture diagram

Simple version:

Video → FFMPEG → Tensorflow/Tesseract/Custom → Callback

Annotated version:

                                 metadata,
                                 progress,
                                  errors
                                    ▲
                                    │
                            ┌───────┴───┐   ┌────────────┐
                         ┌──► stderr    │ ┌─► tesseract  ├─────► callback
                         │  └───────────┘ │ └────────────┘
                         │                │
 ┌─────────────┐ ┌───────┴┐ ┌───────────┐ │ ┌────────────┐
 │ Input Video ├─► ffmpeg ├─► stdout    ├─┼─► tensorflow ├─────► callback
 └─────────────┘ └───────┬┘ └───────────┘ │ └────────────┘
  - Video files          │                │
  - Static images        │  ┌───────────┐ │ ┌────────────────┐
  - HTTP URLs            └──► stdin     │ └─► custom trigger ├─► callback
  - Live streams            └───────┬───┘   └────────────────┘
  - Desktop capture                 │
  - Webcam video                    ▼
                               pause/stop
                                commands

 └─────────────┘ └───────────────────────┘ └─────────────────┘   └──────┘
   MEDIA SOURCE        VIDEO DECODING        COMPUTER VISION     CALLBACK

Getting started (Rust)

cargo install hypetrigger
use hypetrigger::{Hypetrigger, SimpleTrigger};

fn main() {
    Hypetrigger::new()
        .test_input()
        .add_trigger(SimpleTrigger::new(|frame| {
            println!("received frame {}: {}x{}",
                frame.frame_num,
                frame.image.width(),
                frame.image.height()
            );
            // Now do whatever you want with it...
        }))
        .run();
}

Getting started (Typescript)

Browser and Node is supported through a WASM compilation of the image preprocessing code with the excellent Photon.js image processing library. After that Tesseract.js is used for the text recognition.

npm add hypetrigger
const videoElem = document.getElementById('video')
const pipeline = new Hypetrigger(videoElem)
  .addTrigger(frame => {
    console.log({ frame })
    // do whatever you want with the frame
  })
  .autoRun()

Limitations

The TS version is not a fully featured port of the Rust library; it is more of a parallel toolkit with a subset of the full functionality.

There are no Tensorflow.js bindings yet, and frames are pulled directly from media sources, eliminating the use of FFMPEG completely.

For more information, see this page in the docs: Using with other languages.

In-depth example (Rust)

This is slightly simplified sample code. It won't immediately compile and work without the right input source and parameters, but it illustrates how to use the API to solve a real-world problem.

Problem statement: Detect when a goal is scored in live video of a World Cup match. ⚽

Cargo.toml

[dependencies]
hypetrigger = { version = "0.2.0", features = ["photon", "tesseract"] }
# enable the `tesseract` feature and its `photon` dependency for image processing
# see the "Native Dependencies" section in `README.md` if you have trouble building

main.rs

use hypetrigger::{Hypetrigger, SimpleTrigger};
use hypetrigger::photon::{Crop, ThresholdFilter};
use hypetrigger::tesseract::{TesseractTrigger, init_tesseract}

fn main() {
    // First, init a Tesseract instance with default params
    let tesseract = init_tesseract(None, None)?;

    // Initialize some state (use an Rc or Arc<Mutex> if needed)
    let mut last_score: Option<u32> = None;

    // Create a trigger that will be used to detect the scoreboard
    let trigger = TesseractTrigger {
        tesseract, // pass in the Tesseract instance

        // Identify the rectangle of the video that contains
        // the scoreboard (probably the bottom-middle of the
        // screen)
        crop: Some(Crop {
            left_percent: 25.0,
            top_percent: 25.0,
            width_percent: 10.0,
            height_percent: 10.0,
        }),

        // Filter the image to black and white
        // based on text color. This preprocessing improves Tesseract's
        // ability to recognize text. You could replace it with
        // your own custom preprocessing, like edge-detection,
        // sharpening, or anything else.
        threshold_filter: Some(ThresholdFilter {
          r: 255,
          g: 255,
          b: 255,
          threshold: 42,
        }),

        // Attach the callback which will run on every frame with the
        // recognized text
        callback: |result| {
          let parsed_score: u32 = result.text.parse();
          if parsed_score.is_err() {
            return Ok(()) // no score detected; continue to next frame
          }

          // Check for different score than last frame
          if last_score.unwrap() == parsed_score.unwrap() {
            println!("A goal was scored!");

            // Do something:
            todo!("celebrate 🎉");
            todo!("tell your friends");
            todo!("record a clip");
            todo!("send a tweet");
            todo!("cut to commercial break");
          }

          // Update state
          last_score = parsed_score;
        },

        // Using this option will pause after every frame,
        // so you can see the effect of your crop + filter settings
        enable_debug_breakpoints: false,
    };

    // Create a pipeline using the input video and your customized trigger
    Hypetrigger::new()
        .input("https://example.com/world-cup-broadcast.m3u8")
        .add_trigger(trigger)
        .run();

    // `run()` will block the main thread until the job completes,
    // but the callback will be invoked in realtime as frames are processed!
}

Native Dependencies

Visual Studio Build Tools

  • Must install "Visual Studio Build Tools 2017" -- current version 15.9.50
  • Must ALSO install "Visual Studio Community 2019" with the following components of "Desktop development with C++" workload:
    • MSVC v142 - VS 2019 C++ x65/x86 build tools
    • C++ CMake tools for Windows
    • C++ ATL for latest v142 build tools

Build tools are required by Cargo, VS 2019 is used to compile & link native dependencies

Tensorflow

Should be installed automatically by Cargo.

Tesseract

Install manually with vcpkg: (Github)

git clone https://github.com/microsoft/vcpkg
cd vcpkg
./bootstrap-vcpkg.bat
./vcpkg integrate install
./vcpkg install leptonica:x64-windows-static-md
./vcpkg install tesseract:x64-windows-static-md

Also install libclang included in the latest LLVM release.

Current version: https://github.com/llvm/llvm-project/releases/download/llvmorg-14.0.6/LLVM-14.0.6-win64.exe

wasm-pack

cargo install wasm-pack

If you get OpenSSL/Perl errors like this:

This perl implementation doesn't produce Windows like paths

Try running once from windows cmd.exe instead of VSCode integrated terminal and/or git bash.

Dependencies

~13–26MB
~473K SLoC